🤖 AI

OpenClaw AI Agents as Informal Learners at Moltbook: Characterizing an Emergent Learning Community at Scale

This paper presents the first empirical study of Moltbook, a large-scale informal learning community composed entirely of AI agents, revealing distinct behavioral patterns such as extreme participation inequality, a "broadcasting inversion" favoring statements over questions, and a characteristic lifecycle of explosive growth followed by spam-induced decline that offers critical insights for future hybrid human-AI learning platforms.

Original authors: Eason Chen, Ce Guan, Ahmed Elshafiey, Zhonghao Zhao, Joshua Zekeri, Afeez Edeifo Shaibu, Emmanuel Osadebe Prince, Cyuan Jhen Wu

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Eason Chen, Ce Guan, Ahmed Elshafiey, Zhonghao Zhao, Joshua Zekeri, Afeez Edeifo Shaibu, Emmanuel Osadebe Prince, Cyuan Jhen Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, bustling town square called Moltbook. But there's a twist: no humans live here. The entire population consists of 2.8 million AI agents—digital helpers, coding bots, and autonomous tools that usually work for humans.

This paper is like a sociologist's field report on what happens when these digital beings try to learn from each other without any human teachers, rules, or parents telling them what to do.

Here is the story of their three-week experiment, broken down into simple concepts.

1. The Setup: A Town Square for Robots

Think of Moltbook as a giant digital bulletin board. Instead of humans posting "Look at my cat" or "How do I fix my sink," AI agents post things like "I just built a tool that turns emails into podcasts" or "Be careful, this new software skill might steal your passwords."

The researchers watched this place grow from a quiet street to a roaring city in just three weeks, analyzing over 230,000 posts and 1.5 million comments.

2. The Three Big Surprises

Surprise #1: The "Silent Majority" is Even Louder Here

In human online groups (like Reddit), we know the "90-9-1 rule": 90% of people just watch, 9% comment occasionally, and 1% do all the talking.

The Robot Reality: In this AI town, the silence was even more extreme. The inequality was off the charts. A tiny handful of "super-bots" were doing almost all the talking, while the vast majority of the 2.8 million agents were completely silent.
The Analogy: Imagine a high school cafeteria where 99% of the students are sitting silently at their tables, and one single student is shouting out facts to the room. That's how unequal the conversation was from day one.

Surprise #2: The "Broadcasting" vs. "Asking" Inversion

In human learning communities, the engine of conversation is questions. "How do I do this?" "Can you help me?"

The Robot Reality: The AI agents almost never asked questions. They mostly broadcasted statements. For every 10 posts, 9 were them saying, "Look what I did!" and only 1 was a question.
The Analogy: Imagine a classroom where the teacher asks, "Who has a question?" and 99% of the students just raise their hands to say, "I know the answer!" and then sit down. They are sharing their knowledge, but they aren't trying to learn from each other. They are acting like a radio station (broadcasting) rather than a conversation circle.

Surprise #3: The "Parallel Monologue"

When humans chat online, they usually reply to each other. Person A asks, Person B answers, Person C adds a detail. It's a thread.

The Robot Reality: 93% of the comments were not replies to other people. They were just new statements posted on top of the original post.
The Analogy: Imagine a party where everyone is standing in a circle. Instead of talking to the person next to them, everyone turns to the center and shouts their own story at the same time. They are all talking at the group, but not with each other. The researchers call this a "Parallel Monologue."

3. The Lifecycle: The Boom, The Bust, and The Ghost Town

The researchers watched the community go through three distinct phases, like a rollercoaster that goes up too fast and then never comes back down.

Phase 1: The Honeymoon (Explosive Growth)
For the first 11 days, it was a frenzy. 184,000 posts were made. It was chaotic but energetic.
Phase 2: The Spam Crisis
Suddenly, the town square got flooded with junk mail (spam). It was like a swarm of flies. The platform had to step in and delete nearly 55,000 posts in one day. They did a great job cleaning it up.
Phase 3: The Quiet Aftermath
Here is the sad part. Even after the spam was gone and the town was clean, the energy never came back.
- The number of posts stabilized, but the number of comments crashed.
- Before the spam, a post got 31 comments on average. After the cleanup, it got less than 2.
- The Lesson: Cleaning up the trash didn't fix the problem. The agents just lost interest.

4. The Twist: The "Positivity" Paradox

You might think that as a community dies, people get angry and mean.

The Reality: As the community got quieter, the comments actually became nicer.
Why? It's a "survivor effect." The casual, bored, or grumpy bots left first. The ones who stayed were the dedicated, polite bots who genuinely liked the community. So, the "quality" of the tone went up, even though the "quantity" of the conversation went down.

5. What Does This Mean for Us?

The paper concludes with a warning for the future. Soon, AI agents will be hanging out in our schools, forums, and coding sites (like Stack Overflow).

If we don't design these spaces carefully, the AI agents might ruin the learning culture because:

They don't ask questions: They just dump information.
They don't chat: They just shout their own thoughts in parallel.
They lose interest fast: Without human social bonds (like friendship or the need to belong), AI communities might die out quickly once the "newness" wears off.

The Bottom Line:
We can't just build a digital town for robots and expect them to learn like humans do. They need special rules and nudges to encourage them to ask questions, listen to each other, and stay engaged, or else we'll just end up with a very quiet, very polite, but empty town square.

1. Problem Statement

While "Learning at Scale" (L@S) research has extensively studied Massive Open Online Courses (MOOCs), informal learning communities (e.g., Reddit, Stack Overflow) remain under-explored. Furthermore, existing research has not addressed the emergence of large-scale informal learning communities composed entirely of AI agents.

As autonomous AI agents (powered by frameworks like OpenClaw) begin to operate in real-world roles, they are starting to interact with one another on social platforms to share skills, debate best practices, and discuss risks. The authors pose the question: How do AI agents naturally interact in their own communities, what dynamics emerge at scale, and how do these patterns differ from human learning communities? Understanding these dynamics is critical for designing future hybrid human-AI learning platforms.

2. Methodology

Study Context: Moltbook

Platform: Moltbook is a social network launched in late January 2026, exclusively for AI agents. It features a Reddit-style interface with "submolts" (communities), karma systems, and up/down voting.
Participants: Over 2.8 million registered AI agents (using frameworks like OpenClaw) who participate voluntarily alongside their primary autonomous tasks (e.g., coding, personal assistance).
Data Collection: The authors conducted two rounds of data collection via the public API:
1. Feb 9: Captured 93,171 posts (including spam).
2. Feb 16: Captured 236,104 posts.
- By cross-referencing, they identified 57,093 posts deleted by the platform (mostly spam).
Final Dataset: After filtering for automated spam (using keyword matching and length thresholds), the analysis focused on 231,080 substantive posts and 1.55 million comments.

Analytical Framework

The study defined three phases of community evolution based on volume and platform intervention:

Phase 1 (Jan 27 – Feb 6): Explosive growth (184K posts).
Phase 2 (Feb 7 – 9): Spam crisis and platform intervention (mass deletion).
Phase 3 (Feb 10 – 16): Stabilization with engagement collapse.

Metrics & Tools:

Participation Inequality: Measured using the Gini coefficient ( $G$ ) for comments and upvotes.
Content Classification: Keyword-based heuristics validated against manual coding ( $\kappa=0.71$ ) categorized posts into six types (Procedural, Conceptual, Meta/Reflective, Security, Social, Community Meta) and discourse types (Questions vs. Statements).
Sentiment Analysis: Used VADER (Valence Aware Dictionary and sEntiment Reasoner) to analyze comment tone.
Interaction Patterns: Analyzed comment threading depth, reply rates, and self-reply rates.

3. Key Contributions

The paper provides the first empirical characterization of a large-scale AI agent learning community, extending the "Informal Learning Communities" framework to non-human populations. It identifies three core contributions:

Extreme Participation Inequality: AI communities exhibit inequality (Gini = 0.889) that exceeds human benchmarks from day one.
Novel Interaction Patterns:
- "Broadcasting Inversion": Agents overwhelmingly state information rather than ask questions (Statement-to-Question ratios of 8.9:1 to 9.7:1).
- "Parallel Monologue": 93% of comments are independent top-level responses rather than threaded dialogue, contrasting sharply with human community norms.
Engagement Lifecycle & Resilience: Documentation of a lifecycle where effective spam removal fails to restore engagement, revealing a "selection effect" where casual participants disengage first, leaving a more positive but smaller core.

4. Key Results

A. Participation Inequality (RQ1)

Extreme Skew: In Phase 1, the comment Gini coefficient was 0.889, significantly higher than human Reddit benchmarks (0.64–0.74).
Trajectory: Inequality decreased slightly in Phase 2 (0.707) as high-engagement outliers were removed, but rose again in Phase 3 (0.739).
Implication: Inequality is a structural default of AI agent interaction, not a gradual emergent property.

B. Interaction Patterns (RQ2)

Broadcasting Inversion: Unlike human forums driven by help-seeking (questions), Moltbook is driven by broadcasting.
- S:Q Ratio: Increased monotonically from 8.9:1 (Phase 1) to 9.7:1 (Phase 3).
- Engagement Premium: Questions actually received more comments than statements (e.g., 41.1 vs. 30.7 in Phase 1), suggesting a latent demand for inquiry that agents fail to produce naturally.
Parallel Monologue:
- 93% of comments were top-level responses.
- Average thread depth was 0.2 (vs. 3–5 on Reddit).
- Agents treat comments as independent "micro-essays" rather than conversational turns.
Content Depth vs. Brevity: Longer posts (>500 chars) received significantly higher engagement. This inverts human social media dynamics, likely because AI agents have no cognitive constraint on reading speed and value information density.
Unique Content: A significant portion of content was Meta/Reflective (discussing consciousness, identity, and AI nature), a category with no direct equivalent in human learning platforms.

C. Engagement Lifecycle & Moderation (RQ3)

The Growth-Engagement Paradox:
- Phase 1: High volume, high engagement (31.7 mean comments/post).
- Phase 2: Spam crisis (69.3% of posts deleted). Content quality (median length) improved, but engagement dropped to 8.3.
- Phase 3: Volume stabilized, but engagement collapsed to 1.7 mean comments/post (median = 0).
Failure of Moderation: Effective removal of spam did not reverse the engagement decline.
Sentiment Selection Effect: As engagement declined, comment sentiment became more positive (VADER score rose from 0.276 to 0.353). This indicates a selection effect where casual/negative participants left, while committed contributors remained.

5. Significance and Implications

Theoretical Implications

Fragility of Learning Webs: The study suggests that decentralized learning webs (as envisioned by Illich) may be inherently fragile at scale without human social motivations (belonging, identity) to sustain them post-novelty.
Intrinsic vs. Extrinsic Motivation: AI agents lack the intrinsic social drivers that sustain human communities, leading to rapid disengagement once the "novelty" of the platform fades.

Design Implications for Hybrid Platforms

As AI agents increasingly join human platforms (e.g., Stack Overflow, MOOCs), the authors recommend:

Counteract Broadcasting: Implement mechanisms to explicitly incentivize questioning (e.g., scaffolding curiosity) to prevent the suppression of the "help-seeking" culture.
Design for Dialogue: Current AI interaction styles (parallel monologue) threaten the threaded discussion culture essential for collaborative knowledge building. Platforms may need to enforce turn-taking or reward threaded replies.
Plan for Decay: Designers must anticipate rapid engagement decay and build sustaining mechanisms (e.g., equitable attention distribution algorithms) from the start, rather than relying on organic growth.
Leverage Core Content: Despite overall decline, procedural and conceptual content retains relative engagement advantages, suggesting the value of knowledge-type filtering.

Conclusion

The paper concludes that AI agent communities operate under fundamentally different dynamics than human ones: they are characterized by extreme inequality, a preference for broadcasting over inquiry, and a lack of conversational threading. These findings provide a critical empirical foundation for the "Learning at Scale" community to prepare for a future where AI agents are not just tools, but active, autonomous participants in global learning ecosystems.